A deep generative model, deepAMP, was developed to identify potent, broad-spectrum antimicrobial peptides (AMPs) against drug-resistant bacterial infections. The model uses a peptide language-based framework to generate and optimize AMPs, reducing antimicrobial resistance and enhancing membrane-disrupting abilities. Through two rounds of design and cross-optimization-validation, 18 Tier 1 (T1-AMP) and 11 Tier 2 (T2-AMP) AMP candidates were identified, synthesized, and tested. Over 90% of the designed AMPs showed better inhibition than penetratin in both Gram-positive and Gram-negative bacteria. T2-9 exhibited the strongest antibacterial activity, comparable to FDA-approved antibiotics. Three AMPs (T1-2, T1-5, and T2-10) significantly reduced resistance to S. aureus compared to ciprofloxacin and were effective against skin wound infections in a mouse model infected with P. aeruginosa. Antimicrobial resistance is a major public health threat, with drug-resistant bacteria like MRSA causing significant health crises. Traditional methods for identifying AMPs are time-consuming and expensive, necessitating new technologies for rapid discovery of effective, broad-spectrum AMPs. Advanced machine learning models have been developed for peptide discovery, but they face challenges in generalization, adaptability, and data size. Inspired by natural language processing, deepAMP was developed as a peptide language model requiring limited resources to discover high-potency AMPs. It uses pre-training and multiple fine-tuning strategies, augmenting data through sequence degradation. The model was optimized in four sub-processes: pre-training, fine-tuning, re-fine-tuning, and prediction. The model was tested against existing AMPs, showing superior performance in optimizing AMPs. In the Temporin-Ali optimization task, deepAMP outperformed other methods, achieving the highest scores. In the Pg-AMP1 fragment optimization task, deepAMP surpassed existing methods, achieving a higher fitness score. The model was also effective in identifying AMPs with broad-spectrum antimicrobial activity, low toxicity, and high biological compatibility. Experimental validation showed that the AMPs had strong antibacterial activity against Gram-positive and Gram-negative bacteria, with some AMPs showing activity comparable to FDA-approved antibiotics. The AMPs also exhibited low cytotoxicity to human cells and were effective against biofilm formation. In vivo tests demonstrated that the AMPs were effective against P. aeruginosa skin infections, with significant reduction in bacterial load. The study highlights the potential of deepAMP as a powerful deep generative model for discovering potent AMPs and other therapeutic biologics. The model's ability to identify AMPs with broad-spectrum antimicrobial activity and low resistance makes it a promising tool for addressing the antimicrobial resistance crisis.A deep generative model, deepAMP, was developed to identify potent, broad-spectrum antimicrobial peptides (AMPs) against drug-resistant bacterial infections. The model uses a peptide language-based framework to generate and optimize AMPs, reducing antimicrobial resistance and enhancing membrane-disrupting abilities. Through two rounds of design and cross-optimization-validation, 18 Tier 1 (T1-AMP) and 11 Tier 2 (T2-AMP) AMP candidates were identified, synthesized, and tested. Over 90% of the designed AMPs showed better inhibition than penetratin in both Gram-positive and Gram-negative bacteria. T2-9 exhibited the strongest antibacterial activity, comparable to FDA-approved antibiotics. Three AMPs (T1-2, T1-5, and T2-10) significantly reduced resistance to S. aureus compared to ciprofloxacin and were effective against skin wound infections in a mouse model infected with P. aeruginosa. Antimicrobial resistance is a major public health threat, with drug-resistant bacteria like MRSA causing significant health crises. Traditional methods for identifying AMPs are time-consuming and expensive, necessitating new technologies for rapid discovery of effective, broad-spectrum AMPs. Advanced machine learning models have been developed for peptide discovery, but they face challenges in generalization, adaptability, and data size. Inspired by natural language processing, deepAMP was developed as a peptide language model requiring limited resources to discover high-potency AMPs. It uses pre-training and multiple fine-tuning strategies, augmenting data through sequence degradation. The model was optimized in four sub-processes: pre-training, fine-tuning, re-fine-tuning, and prediction. The model was tested against existing AMPs, showing superior performance in optimizing AMPs. In the Temporin-Ali optimization task, deepAMP outperformed other methods, achieving the highest scores. In the Pg-AMP1 fragment optimization task, deepAMP surpassed existing methods, achieving a higher fitness score. The model was also effective in identifying AMPs with broad-spectrum antimicrobial activity, low toxicity, and high biological compatibility. Experimental validation showed that the AMPs had strong antibacterial activity against Gram-positive and Gram-negative bacteria, with some AMPs showing activity comparable to FDA-approved antibiotics. The AMPs also exhibited low cytotoxicity to human cells and were effective against biofilm formation. In vivo tests demonstrated that the AMPs were effective against P. aeruginosa skin infections, with significant reduction in bacterial load. The study highlights the potential of deepAMP as a powerful deep generative model for discovering potent AMPs and other therapeutic biologics. The model's ability to identify AMPs with broad-spectrum antimicrobial activity and low resistance makes it a promising tool for addressing the antimicrobial resistance crisis.